Cross-Boundary Projections Diagnostics

Model Evaluation for Cross Boundary

The following sections will cover how to pull out common diagnostics from single-species SDMTMB model outputs and some data visualization development for cross boundary metrics for US and Canadian population densities.

Single Species Projection Results

Andrew has been processing individual species distribution models for a suite of species that inhabit the Northeast US and Scotian shelf habitate. These models are developed using a combined catch dataset built from the US and Canadian fisheries independent trawl survey programs.

There are three model “types” being explored to investigate different model structures and their tradeoffs when applied for projected species distributions.

Loading sdmTMB Models

The models themselves provide details on the formula’s used and can be used to produce marginal effects plots of the species preferences for depth and temperature.

Load Model Projections

The projections have been run for two SSP scenarios, which extend forward in time through 2100.

Plotting Preference Curves

Preference curves can be produced by providing a prediction dataframe to the predict() function/method. For this, I set provide a range of values for the effecto of interest and a value of 0 for the other fixed effects. The predictors are scaled so this process holds the other effects at the mean value while exploring a range of values for the one we’re interested in.

Spatially Varying Season Effects

In all three models, the seasonal effect is fit as a spatially explicit surface.

SeasonSVC

Seasonal Marginal Effects

Seasonal Zetas

Season Zetas + Main Effect

Model Type Impact on Seaonal Effects

Baseline Period Maps

For these maps a baseline period of the most recent 20 years is used (2004-2023) as a benchmark for comparison.

Assign Nationality Labels to Projections

For cross-boundary we are interested in differences between US and Canadian study areas. This section takes the unique projection locations and assigns the national jurisdiction that each location falls within.

This chunk of code applies the approach for an approachable subset of data, and makes a map of the baseline biomass estimates.

Baseline Period Map

This is the full comparison matrix:

Mid-Century Biomass Map

Mid-Century Change

End of Century Biomass

End of Century Change

Overall Biomass Changes

See the biomass changes (overall, US, Can) as time series over years for the outputs of different model structures and forward-projection methods within the s-t structure

For the whole USA + Canada Region, here is what the timeseries for a single species’ projected biomass could resemble

Territory Timeseries

Within each national jurisdiction (USA, CAN) we can estimate the total biomass from the predicted densities from the model.

Territory Proportions

What side of the hague line is most of the biomass located within?

Center of Gravity

  1. See the COG changes (across whole domain) as a time series over years for the outputs of different model structures and forward-projection methods within the s-t structure

Modeling Choice Diagnostics

Get outputs from Andrew for cod, halibut, and lobster and make tables or visuals to help us:

Forward Projection of Error:

  1. See the effects of different forward-projection methods within spatio-temporal model (e.g., AR1 vs RW vs IID year-season selection) i.e., how does the choice of method affect resulting trends in biomass, COG, variance, etc.?

Model Performance Statistics

  1. Compare model performance statistics (AIC, explanatory power, RMSE, etc.) for different model structures (i.e., null/env’t-only, spatial, spatio-temporal)–I’m thinking this could just be a table

Regrouping

  1. I’m thinking this would be a relatively large set of figures and tables that include the following for all model structures (i.e., null, spatial, spatio-temporal AR1, s-t RW, s-t IID) and two climate scenarios (“overall”=over whole domain):
  2. Biomass overall maps (baseline, mid-century, end of century)
  3. Biomass overall time series with variance
  4. Biomass US vs Can time series with variance
  5. COG overall time series

Model performance stats:

If you have other ideas of things we should look at / ways to look at this to help us get to final model selection, please let us know.